Many types of semiconductor devices are made using similar manufacturing procedures. A starting substrate, usually a thin wafer of silicon, is doped, masked, and etched through several process steps, the steps depending on the type of devices being manufactured. This process yields a number of die on each wafer produced. Each die on the wafer is given a brief test for full functionality, and the nonfunctional die are mechanically marked or mapped in software. This brief test is only a gross measure of functionality, and does not insure that a die is completely functional or has specifications that would warrant its assembly in a package.
If the wafer has a yield of grossly functional die which indicates that a good quantity of die from the wafer are likely to be fully operative, the die are separated with a die saw, and the nonfunctional die are scrapped while the rest are individually encapsulated in plastic packages or mounted in ceramic packages with one die in each package. After the die are packaged they are rigorously electrically tested. Components which turn out to be nonfunctional or which operate at questionable specifications are scrapped or devoted to special uses.
Packaging unusable die only to scrap them after testing is a waste of time and materials, and is therefore costly. Given the relatively low profit margins of commodity semiconductor components such as dynamic random access memories (DRAMs) and static random access memories (SRAMs), this practice is uneconomical. However, no thorough and cost effective method of testing an unpackaged die is available which would prevent this unnecessary packaging of nonfunctional and marginally functional die.
It is proposed that multiple integrated circuit devices be packaged as a single unit, known as a multi chip module (MCM). This can be accomplished with or without conventional lead frames. This creates two problems compared to conventional test methods. Firstly, discrete testing is more difficult because the conventional lead frame package is not used. Furthermore, when multiple devices are assembled into a single package, the performance of the package is reduced to that of the die with the lowest performance. In other words, the ability to presort the individual dice is limited to that obtained through probe testing. Secondly, the packaging may have other limitations which are aggravated by burn-in stress conditions so that the packaging becomes a limitation for burn-in testing.
The practice of packaging die only to find the component must be scrapped can especially adversely affect yields on multi-chip modules (MCMs). With MCMs, several unpackaged die are assembled into a single component, then the component is tested as a single functional unit. If a single die is nonfunctional or operates outside of acceptable specifications, the entire component fails and all die in the package are scrapped or an attempt is made to "re-work" the MCM. There is presently no cost-effective way to reclaim the functioning die. Statistically, the yields of MCMs decrease in proportion to the increasing number of die in each module. The highest density modules have the lowest yields due to their increased total silicon surface area. For discretely packaged parts, if the product yield of good parts from preliminary testing to final shipment (probe-to-ship) is, for example, 95%, one would not be particularly concerned with packaging costs for the failed parts, if packaging costs are 10% of the product manufacturing costs. Where packaging costs are considerably higher, however, as in ceramic packaged parts, testing a die before packaging is economical when the cost of packaging divided by the package yield is equal to or greater than the cost of testing: EQU C.sub.DIE .times.(C.sub.PACKAGE /Package Yield)=C.sub.DIE .times.C.sub.ADDL.KGD
where
C=cost PA1 C.sub.DIE =manufacturing cost of functional site PA1 C.sub.ADDL.KGD =additional cost of testing unpackaged die in order to produce known good die (KGD)
Note that in the case of discretely packaged parts, the cost of the die (C.sub.DIE) is essentially not a factor since it is the same on both sides of the equation. This changes in the case of MCMs having more than one part type, for example memory and a microprocessor. Scrapping (or reclaiming) the microprocessor after packaging because of malfunctioning memory is much more costly than scrapping a module containing only memory: EQU (C.sub.DIE).times.(Number of Die/Die Yield).times.C.sub.PACKAGE =C.sub.DIE .times.C.sub.ADDL.KGD
The above equation must be modified to account for varied costs and yields of die in modules with mixed part types. With MCMs, the cost of packaging a failed part is proportional to the number of die in the module. In the case of a memory module having 16 die, where probe-to-ship yield of the die is 95%, the costs are: EQU 16/0.95.times.C.sub.PACKAGE =C.sub.ADDL.KGD
so the additional costs of testing for known good die (KGD) may be 16 times the cost of testing an unrepairable module and still be economical. This, of course, is modified by the ability to repair failed modules.
Testing of unpackaged die before packaging would be desirable as it would result in reduced material waste, increased profits, and increased throughput. Using only known good die in multichip modules would increase yields significantly.